METHOD FOR GENERATING DESIGN DATA INCLUDING SUBSTRUCTURE SELECTION STEP, METHOD FOR GENERATING DESIGN DATA FOR FUNCTIONAL MOLECULES, DESIGN PROGRAM FOR FUNCTIONAL MOLECULES, AND FUNCTIONAL MOLECULES

Abstract

To provide a method and a design program for generating design data of a substance having promising properties, by utilizing machine learning to generate design data of the substance, and selecting a promising partial structure by paying attention to enrichment of the partial structure, and to provide a molecule which has been designed according to the method.SOLUTION: A method for generating design data of a substance having properties that are predictable using computer simulation includes: a first generation step S1 for generating first generation data according to machine learning using training data; a second generation step S2 for generating second generation data by screening substances having properties, by using computer simulation, from the first generation data; a partial structure selection step S3 for selecting a partial structure on the basis of enrichment of the partial structure of the substance, from the second generation data; and a third generation step S4 for generating design data of the substance by screening substances having the partial structure selected in the partial structure step, from the second generation data.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a method for generating design data including a substructure selection step, a method for generating design data for a functional molecule, a design program for a functional molecule, and a functional molecule. One embodiment of the present invention relates to a method for generating design data for an organic molecule having long-wavelength light absorption ability, a design program for an organic molecule having long-wavelength light absorption ability, and an organic molecule having long-wavelength light absorption ability. [Background technology]

[0002] For a long time, attention has been focused on technologies that allow computers to design molecules with desired functions. However, most of these require humans to input the chemical rules that govern molecular composition in advance, which is labor-intensive and impossible to cover all rules. However, in recent years, advances in artificial intelligence (AI) technology, from machine learning to deep learning, have made it possible to automatically teach computers the rules governing the composition of complex organic molecules. This has led to dramatic advances in the technology of using AI to design functional molecules, and many new molecules have been designed.

[0003] Monte Carlo Tree Search (MCTS) is one of the algorithms often used in reinforcement learning, a type of machine learning. Monte Carlo Tree Search determines the best action to take in the current state by representing the search space as an incrementally growing tree structure (search tree). The search tree consists of nodes that represent states and edges that represent actions. Each node holds the expected reward for moving forward from that node.

[0004] Monte Carlo tree search is divided into four steps: selection, expansion, simulation (roll-out), and backpropagation. In the selection step, child node selection is repeated from the root node until a node that is not yet fully expanded is reached, and then until a leaf node is reached. In the expansion step, a node is added to the leaf node selected in the selection step. In the simulation step, a simulation is performed according to the default roll-out policy and a reward is obtained for the selected node. In the backpropagation step, the results of the simulation are propagated in the reverse direction, from the leaf node to the root node. As mentioned above, nodes represent the states of the simulated task, and edges correspond to the actions performed to transition from the current node state to the child node state. The expected reward of a node represents the average outcome of all simulations that proceed through that node.

[0005] As an example of using Monte Carlo tree search in molecular design, Non-Patent Document 1 proposed ChemTS, a molecule generator that combines a recurrent neural network (RNN) and Monte Carlo tree search to discover new and useful molecules. In ChemTS, molecules are represented as strings in the Simplified Molecular Input Line Entry System (SMILES) notation, and each element of the SMILES string corresponds to a node in the search tree of the Monte Carlo tree search. To generate realistic molecules in steps 2 and 3 of the four steps above, an RNN-based SMILES character prediction model is used from incomplete SMILES strings. In order to design molecules with desired properties, it is necessary to evaluate the molecules generated in step 3 using appropriate criteria. As disclosed in Non-Patent Document 1, we have shown that by setting a (black box) evaluation function of the J score including the logP value in step 3, it is possible to efficiently generate designed molecules with high logP values.

[0006] On the other hand, in Non-Patent Document 2, ChemTS and density functional theory (DFT) were combined to design molecules that absorb the target wavelength, and the possibility of automatic molecular design was experimentally verified. [Prior art documents] [Non-patent literature]

[0007] [Non-Patent Document 1] Yang, X.; Zhang, J.; Yoshizoe, K.; Terayama, K.; Tsuda, K. ChemTS: an efficient python library for de novo molecular generation. Sci. Technol. Adv. Mater. 2017, 18, 972-976. [Non-patent document 2] Sumita, M.; Yang, X.; Ishihara, S.; Tamura, R.; Tsuda, K. Hunting for Organic Molecules with Artificial Intelligence: Molecules Optimized for Desired Excitation Energies. ACS Cent. Sci. 2018, 4, 1126-1133. Summary of the Invention [Problem to be solved by the invention]

[0008] However, when there are a large number of molecular structures directly output by the methods disclosed in Non-Patent Documents 1 and 2, there is a problem in that it is time-consuming and expensive to actually synthesize and evaluate all of the molecules. There has been a demand for the selection of more effective functional molecules from the molecules output by a machine learning method that combines ChemTS calculations and density functional theory (DFT) calculations to find functional molecules with predetermined physical properties.

[0009] The present invention has been made to solve the above-mentioned problems, and aims to provide a method and a design program for screening promising objects and generating object design data by selecting promising partial structures from object design data generated directly by machine learning or the like, focusing on the enrichment of the partial structures. Another object of the present invention is to provide a method and design program for screening promising molecules from molecular data generated directly by machine learning or the like, and generating design data for functional molecules. Another object of the present invention is to provide a molecule designed by a method for generating design data for a functional molecule. [Means for solving the problem]

[0010] The present invention includes the following aspects. [1] A method for generating design data for an object having properties that can be predicted by computer simulation, comprising: a first generation step of generating first generated data by machine learning using training data; a second generation step of generating second generation data by screening the object having the property from the first generation data through computer simulation; a substructure selection step of selecting a substructure from the second generated data based on the enrichment of the substructure of the object; a third generation step of screening the object having the substructure selected in the substructure selection step from the second generation data to generate design data of the object; Including, A method for generating design data, wherein the enrichment of the partial structure is a value calculated using the following formula (1A): Enrichment = Proportion of objects containing the substructure in the second generation data / Proportion of objects containing the substructure in the training data (1A) [2] A method for generating design data for a functional molecule having physical properties that can be calculated by quantum chemical calculation, comprising: a first generation step of generating first generated data by machine learning using training data; a second generation step of generating second generation data by screening the functional molecule from the first generation data using quantum chemical calculations; a substructure selection step of selecting a substructure that is a part of the functional molecule from the second generated data using an enrichment of the substructure; a second generation step of screening functional molecules having the partial structure from the second generation data to generate second generation data; Including, A method for generating design data of a functional molecule, wherein the enrichment of the partial structure is a value calculated by the following formula (1B): Enrichment = fraction of molecules containing the substructure in the second generation data / fraction of molecules containing the substructure in the training data (1B) [3] The physical property is a light absorption wavelength, The method for generating design data of a functional molecule according to [2], wherein the functional molecule is an organic molecule capable of absorbing long-wavelength light. [4] The machine learning is reinforcement learning using a Monte Carlo Tree Search (MCTS) algorithm, The method for generating design data for functional molecules according to [2] or [3], wherein in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network (RNN) and quantum chemical calculations. [5] The Monte Carlo tree search includes a selection procedure, an expansion procedure, a simulation procedure, and a backpropagation procedure; In the selection procedure, the score u of the child node of the parent node is calculated using the following formula (2): i is calculated using the UCB1 (Upper Confidence Bound 1) algorithm,

number

number

[10] A program for causing a computer to execute a method for generating design data for a functional molecule having physical properties that can be calculated by quantum chemical calculation, the method comprising: a first generation step of generating first generated data by machine learning using training data; a second generation step of generating second generation data by screening the functional molecule from the first generation data using quantum chemical calculations; a substructure selection step of selecting a substructure that is a part of the functional molecule from the second generated data using an enrichment of the substructure; a second generation step of screening functional molecules having the partial structure from the second generation data to generate second generation data; Including, The program is characterized in that the enrichment of the partial structure is a value calculated using the following formula (1B): Enrichment = fraction of molecules containing the substructure in the second generation data / fraction of molecules containing the substructure in the training data (1B)

[11] The physical property is a light absorption wavelength, The program according to

[10] , wherein the functional molecule is an organic molecule capable of absorbing long-wavelength light.

[12] A computer-readable recording medium having a program recorded thereon for causing a computer to execute a method for generating design data for an object having properties that can be predicted by computer simulation, the method comprising: a first generation step of generating first generated data by machine learning using training data; a second generation step of generating second generation data by screening the object having the property from the first generation data through computer simulation; a substructure selection step of selecting a substructure from the second generated data based on the enrichment of the substructure of the object; a third generation step of screening the object having the substructure selected in the substructure selection step from the second generation data to generate design data of the object; Including, The recording medium, wherein the enrichment degree of the partial structure is a value calculated by the following formula (1A): Enrichment = Proportion of objects containing the substructure in the second generation data / Proportion of objects containing the substructure in the training data (1A)

[13] A computer-readable recording medium storing a program for causing a computer to execute a method for generating design data for functional molecules having physical properties that can be calculated by quantum chemical calculation, the method comprising: a first generation step of generating first generated data by machine learning using training data; a second generation step of generating second generation data by screening the functional molecule from the first generation data using quantum chemical calculations; a substructure selection step of selecting a substructure that is a part of the functional molecule from the second generated data using an enrichment of the substructure; a second generation step of screening functional molecules having the partial structure from the second generation data to generate second generation data; Including, The recording medium, wherein the enrichment degree of the partial structure is a value calculated by the following formula (1B). Enrichment = fraction of molecules containing the substructure in the second generation data / fraction of molecules containing the substructure in the training data (1B)

[14] The physical property is a light absorption wavelength, A computer-readable recording medium having the program according to

[13] recorded thereon, wherein the functional molecule is an organic molecule capable of absorbing long-wavelength light.

[15] Molecules represented by the following formulas (5-1) to (5-22) having a 1,2-naphthoquinone structure. [ka]

[16] A molecule represented by the following formula (6) or a derivative thereof: [ka] [Effects of the Invention]

[0011] According to the present invention, it is possible to provide a method and a design program for generating design data of an object by using machine learning and selecting promising substructures by focusing on the enrichment of the substructures, thereby screening promising objects and generating design data of the object. According to one embodiment of the present invention, there are provided a method for generating design data for a functional molecule, which generates molecules using machine learning and focuses on the enrichment of substructures that contribute to predetermined physical properties, thereby enabling the design of promising functional molecules. Furthermore, there is also provided a functional molecule obtained by the method for generating design data for a functional molecule. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram for explaining a first embodiment. [Figure 2] FIG. 10 is a diagram for explaining a second embodiment. [Figure 3] FIG. 10 is a diagram for explaining an example of a second embodiment. [Figure 4] FIG. 1 is a diagram for explaining a molecular search using Monte Carlo tree search (MCTS). [Figure 5] FIG. 1 shows the design of molecule 1d that can be synthesized from molecule (5-17) (molecule 1a) included in the second generation data obtained in Example 1. [Figure 6] FIG. 1 shows a retrosynthetic analysis of a synthesizable molecule 1d designed from molecule (5-17) (molecule 1a) included in the second generation data obtained in Example 1. [Figure 7] FIG. 1 shows a synthetic scheme for molecule 1d. [Figure 8] The UV-visible absorption spectrum of molecule 1d in acetonitrile (10-4 mol / L) and the calculated spectrum of molecule 1d obtained by TD-DFT calculation (calculation level: APFD / 6-311+++G**) are shown. The red line (solid line) shows the UV-visible absorption spectrum, and the blue line (dotted line) shows the calculated spectrum. DETAILED DESCRIPTION OF THE INVENTION

[0013] Hereinafter, embodiments of the present invention will be described in detail, but the present invention is not limited to the embodiments described below.

[0014] The terms of the present invention have the following meanings. "Machine learning": A type of artificial intelligence, or a field of study, that describes computer algorithms that automatically improve through learning. They learn using data called "training data" or "learning data," and then use the learning results to accomplish a task.

[0015] "Training data": Data used to train a model in machine learning. In the present invention, this also includes data used to train the recurrent neural network described below.

[0016] "Recurrent Neural Network (RNN)": A type of artificial neural network in which connections between nodes form a directed graph along a sequence. This allows it to show dynamic temporal behavior for time series. In the present invention, for example, after training, it is one of the deep learning methods that can predict the next character from the character that appeared immediately before.

[0017] "Deep Learning": A machine learning technique that combines neural networks in multiple layers to enhance their representation and learning capabilities.

[0018] Monte Carlo Tree Search (MCTS) is one of the most effective search methods for games such as Go. It is a tree search method using the Monte Carlo method. It is a heuristic search algorithm for the decision process (i.e., it stops unnecessary searches midway and can lead to good moves with a fairly high probability).

[0019] "Computer simulation": Includes simulations based on computational science and simulations based on learning models generated by machine learning.

[0020] "Quantum Chemistry": A molecular simulation technology based on quantum mechanics. Quantum mechanics is a type of mechanics that explains particle phenomena at the atomic and molecular level.

[0021] "Density Functional Theory (DFT)" is a simulation based on quantum mechanics for obtaining the electronic state of molecules and materials, and is one of the calculation methods in quantum chemistry.

[0022] "Screening": Refers to the process of selecting from multiple targets those that meet certain conditions.

[0023] {First embodiment} (Method for generating design data) The method for generating design data according to this embodiment is a method for generating design data for an object having properties that can be predicted by computer simulation. As shown in Fig. 1, the method includes the following first generation step (S1), second generation step (S2), substructure selection step (S3), and third generation step (S4). First generation step (S1): A step of generating first generation data using training data through machine learning. Second generation step (S2): A second generation step (S2) of generating second generation data by screening, from the first generation data, entities having the properties described above through computer simulation. Substructure selection step (S3): A substructure selection step (S3) of selecting a substructure from the second generated data based on the enrichment of the substructure of the object. Third generation step (S4): A step of screening the object having the partial structure selected in the partial structure selection step (S3) from the second generation data, and generating design data for the object.

[0024] The enrichment degree is expressed by the following formula (1A).

[0025] Enrichment = Proportion of items containing the substructure in the second generation data / Proportion of items containing the substructure in the training data (1A)

[0026] The object of this embodiment is not particularly limited as long as it has properties that can be predicted by computer simulation. Examples of the object of this embodiment include inorganic molecules, organic molecules, composite materials, ceramic materials, and metal materials in the chemical and material fields. Examples of the object of this embodiment include electric circuits, electric devices, and magnetic sensors in the electrical and magnetic fields. Examples of the object of this embodiment include musical compositions, images, photographs, and film works in the music and video fields.

[0027] The "property that can be predicted by computer simulation" in this embodiment is not particularly limited as long as it is a property that is determined by the configuration of the object and can be predicted by computer simulation using data on the configuration of the object. Examples of the "property" include the inherent objective characteristics of the object and subjective evaluations of the object. In this embodiment, the "properties that can be predicted by computer simulation" include, for example, the physical properties of chemical molecules or materials, such as absorption wavelength, dielectric properties, redox potential, polarization properties, ionic conductivity, and cohesive energy, which will be described in the second embodiment below. In this embodiment, the "properties that can be predicted by computer simulation" include, for example, electrical characteristics of the output of an electric circuit, such as current, voltage, and frequency. In this embodiment, the "properties that can be predicted by computer simulation" include, for example, when the object is a musical composition, a ranking that evaluates the popularity of the musical composition, a hit rate, and the like.

[0028] <First generation step (S1)> The computer simulation of this embodiment is not particularly limited as long as it can predict predetermined properties using the structure of the object on a computer. For example, various simulation methods reported in computational science can be used. For example, when the object is a chemical molecule or material, quantum chemical calculations described in the second embodiment below can be used. The computer simulation of this embodiment includes not only simulation methods such as quantum chemical calculations used in computational science, but also machine learning methods using a predictive learning model. For example, a learning model can be generated by a machine learning method using learning data of the object having the properties, and the properties can be predicted from design data of the object using the learning model.

[0029] The training data (sometimes referred to as teacher data, learning data, or training data) in this embodiment is, for example, data on at least the structure of an object, and is data used for training. Data containing data on the structure of an object having a predetermined property and data on the property of the object can also be used as training data. For example, if the entity is a molecule, the training data for the molecule may be data on the chemical structure of the molecule, as will be described in detail in the second embodiment below.

[0030] The machine learning according to this embodiment is not particularly limited as long as it can design an object having predetermined properties using the training data, and examples thereof include supervised learning, unsupervised learning, and reinforcement learning. In a second embodiment described later, an example using reinforcement learning such as Monte Carlo Tree Search (MCTS) will be described. The machine learning according to this embodiment preferably uses a neural network, and more preferably uses deep learning. In the embodiment described below, an example using a recurrent neural network (RNN) will be described. In this embodiment, the training data is data for training the recurrent neural network (RNN).

[0031] The machine learning according to the present embodiment uses reinforcement learning such as Monte Carlo tree search, and in the Monte Carlo tree search, the expansion and rollout are preferably performed by combining a recurrent neural network and computer simulation. More preferably, the computer simulation is used to calculate the value of the property from the leaf node for the configuration obtained by the recurrent neural network trained with the training data. When designing a product with better properties, the higher the property value, for example, when designing an organic molecule with a long wavelength (described below), it is more preferable to use the calculated property value and set the reward for the Monte Carlo tree search so that the larger the property value is above a predetermined lower limit. Alternatively, when designing a product with better properties, the lower the property value is, for example, when designing a molecule with a low dielectric constant, it is more preferable to use the calculated value and set the reward for the Monte Carlo tree search so that the smaller the property value is below a predetermined upper limit. By setting up the machine learning in this way, it is possible to design new products that exceed the limitations of the products included in the training data.

[0032] <Second generation step (S2)> In the second generation step (S2), the first generation data obtained in the first generation step (S1) includes configuration data of the objects and value data of properties obtained by computer simulation, and the objects having certain values ​​are screened from the first generation data to generate second generation data. When designing a substance with better properties, the higher the value, the better the properties, the more the substances with a predetermined value or higher are screened, and the second generated data is generated. For example, when designing an organic molecule with a long wavelength (described below), organic molecules with a predetermined light absorption wavelength or higher are screened, and the second generated data is generated. Alternatively, when designing a substance with better properties, the lower the value, the better the properties, the more the substances with a predetermined value or lower are screened, and the second generated data is generated. For example, when designing a molecule with a low dielectric constant, molecules with a predetermined dielectric constant or lower are screened, and the second generated data is generated.

[0033] In the second generation step (S2), the property values ​​of the first generation data may be recalculated using only the structure of the objects in the first generation data obtained from the first generation step (S1) using a simulation method different from that used in the first generation step (S1) or the same computer simulation but with different calculation accuracy.The recalculated values ​​may then be used to screen the first generation data for objects having certain values ​​to generate second generation data.This allows for both a high speed first generation step (S1) and high accuracy in the second generation step (S2).

[0034] In the second generation step (S2), the computer simulation of the first generation data obtained in the first generation step (S1) and data on property values ​​may be used to first screen the objects having a predetermined first value to generate intermediate data, and the property values ​​of the objects in the intermediate data may be recalculated using a simulation method different from that used in the first generation step (S1) or the same computer simulation but with a different calculation accuracy, and then the recalculated property values ​​may be used to screen the objects having a predetermined second value from the intermediate data to generate second generation data.

[0035] <Substructure selection step (S3)> The "substructure" mentioned above is a part of the structure contained in the substance that particularly contributes to the above properties.

[0036] One or more of the partial structures may be selected from the partial structure candidate list. There are no particular limitations on the method for selecting a partial structure that contributes to the above properties, and the following three methods can be mentioned, for example. [1] A method of selecting a partial structure based on the properties and common technical knowledge about the objects contained in the training data; [2] A method for automatically extracting and selecting possible substructures by analyzing the objects and their properties contained in the training data; [3] A method of using a database of partial structures and selecting those contained in the database that contribute to the above properties. In the examples described below, the eight partial structures shown in Table 1 were selected from the viewpoint of contributing to long wavelength light absorption, and a partial structure candidate list for the embodiment was created.

[0037] For the selected partial structure candidate list, the respective proportions are calculated in the training data and the second generated data obtained in the second generation step (S2), and the enrichment of each partial structure is calculated using the formula (1A). For example, if the number of entities contained in the training data is Nt and the number of entities containing substructure f is Ntf, then the proportion of entities containing substructure f in the training data, Pt(f) = Ntf / Nt. Similarly, if the number of entities contained in the second generated data is Nd and the number of entities containing substructure f is Ndf, then the proportion of entities containing substructure f in the second generated data, Pd(f) = Ndf / Nd. Using formula (1A), Pd(f) / Pt(f) is calculated as the enrichment of substructure f.

[0038] In the second generated data, a partial structure is selected using enrichment. A partial structure with a high enrichment can be selected. This is because entities having a partial structure with a high enrichment are particularly important entities among entities generated by machine learning. For example, partial structures with the first to third highest enrichments can be selected in descending order of enrichment. It is preferable to select the partial structure with the highest enrichment. In the examples described below, the partial structure with the highest enrichment was selected from the partial structure candidate list in Table 1 of the examples.

[0039] <Third generation step (S4)> Then, from among the objects contained in the second generated data, objects containing the partial structure selected in the partial structure selection step (S3) are screened, and design data for an object having predetermined properties is generated. In the examples described below, organic molecule design data was generated that included 22 molecules with the most highly enriched partial structures from the molecules included in the second generated data.

[0040] <Optimal design step> The method of generating design data of a functional molecule of this embodiment may further include an optimum design step of designing an optimum product from the design data obtained in the third generation step (S4). The optimization design step may, for example, involve re-evaluating the design data obtained in the third generation step (S4) using a different computer simulation from that used in the previous step, or re-evaluating the design data using the same computer simulation method as that used in the previous step but with a higher degree of accuracy (which requires more computational time). The optimization design step may, for example, involve actually manufacturing an object contained in the design data obtained in the third generation step (S4) or an object having a similar configuration, and evaluating the properties of the manufactured object. A product having a similar configuration to that contained in the design data is a product that, from the perspective of the configuration that contributes to the properties, has the same basic configuration as the object contained in the design data, but has a configuration in which partial structures that are difficult to manufacture are removed or replaced with partial structures that are easier to manufacture.

[0041] (program) The design program of this embodiment is a program for causing a computer to execute a method for generating design data for an object having properties that can be predicted by computer simulation. A first generation step (S1) of generating first generated data by machine learning using training data; a second generation step (S2) of generating second generation data by screening the object having the property from the first generation data through computer simulation; a substructure selection step (S3) of selecting a substructure from the second generated data based on the enrichment of the substructure of the object; a third generation step (S4) of screening the object having the substructure selected in the substructure selection step from the second generation data to generate design data of the object; Includes. The enrichment degree of the partial structure is a value calculated by the following formula (1A).

[0042] Enrichment = Proportion of items containing the substructure in the second generation data / Proportion of items containing the substructure in the training data (1A)

[0043] In the above-described method for generating design data, a preferred embodiment of the method for generating design data is described. From another viewpoint, the program of this embodiment is a program for causing a computer to execute each step of the preferred embodiment of the method for generating design data.

[0044] (Computer-readable recording medium on which a program is recorded) The present invention also includes a computer-readable recording medium having recorded thereon a program for generating design data for the above-mentioned product and a preferred embodiment. Examples of the recording medium include tape-based media such as magnetic tape and cassette tape, disk-based media including magnetic disks such as floppy disks (registered trademark) and hard disks, and optical disks such as CD-ROMs, MOs, MDs, DVDs, and CD-Rs, card-based media such as IC cards (including memory cards) and optical cards, and semiconductor memory-based media such as mask ROMs, EPROMs, EEPROMs (registered trademark), and flash ROMs.

[0045] {Second embodiment} (Method for generating design data for functional molecules) The method for generating design data for functional molecules in this embodiment includes the following first generation step (S11), second generation step (S12), substructure selection step (S13), and third generation step (S14), as shown in Figure 2. First generation step (S11): A step of generating first generation data by machine learning using training data. Second generation step (S12): A second generation step (S12) of generating second generation data by screening the functional molecules from the first generation data using quantum chemical calculations. Substructure selection step (S13): A substructure selection step (S13) of selecting a substructure that is a part of the functional molecule from the second generated data using the enrichment of the substructure. Third generation step (S14): A step of screening molecules having the partial structure from the second generation data to generate design data. The functional molecule of this embodiment is a molecule having physical properties that can be calculated by quantum chemical calculations, and the physical properties are preferably at least one selected from, for example, absorption wavelength, dielectric properties, redox potential, polarization properties, ionic conductivity, and cohesive energy. The enrichment degree is expressed by the following formula (1B).

[0046] Enrichment = fraction of molecules containing the substructure in the second generation data / fraction of molecules containing the substructure in the training data (1B)

[0047] <First generation step (S11)> Quantum chemistry is a molecular simulation technology based on quantum mechanics, which is a type of mechanics that explains particle phenomena at the atomic and molecular level. In a third embodiment described below, the optical absorption wavelength of a designed organic molecule is calculated by obtaining the electronic state of the organic molecule through quantum chemical calculations based on density functional theory. However, the specific quantum chemistry theory to be used can be appropriately selected depending on the physical properties, the computational power of the computer used, the limited computation time, etc.

[0048] Training data (sometimes called teacher data, learning data, or training data) is, for example, data on at least the chemical structure of a compound, and is data used to train a model. Data on the chemical structure of a compound having predetermined physical properties and data on the physical properties of the compound can also be used as training data. The data on the chemical structure of a molecule is preferably written in a linear notation such as SMILES notation, SMARTS notation, InChl notation, etc. In the examples described below, an example using SMILES notation will be shown.

[0049] The machine learning according to this embodiment is not particularly limited as long as it can design molecules having predetermined physical properties using the training data. For example, supervised learning, unsupervised learning, and reinforcement learning can be used. In the embodiment described below, an example using reinforcement learning such as Monte Carlo Tree Search (MCTS) will be described. The machine learning according to this embodiment preferably uses a neural network, and more preferably uses deep learning. In the embodiment described below, an example using a recurrent neural network (RNN) will be described. In this embodiment, the training data is data for training the recurrent neural network (RNN) (e.g., S31 and S32 in FIG. 3).

[0050] The machine learning according to this embodiment uses reinforcement learning such as Monte Carlo tree search, as shown in Fig. 3. In the Monte Carlo tree search, the expansion and rollout are preferably performed by combining a recurrent neural network and quantum chemical calculations such as density functional theory. It is more preferable to calculate the physical property values ​​of the designed molecules obtained from the leaf nodes by quantum chemical calculations such as density functional theory using the recurrent neural network trained with the training data. When designing a functional molecule with better physical properties, the higher the physical property value, for example, when designing an organic molecule with a long wavelength (described below), it is more preferable to set the calculated physical property value so that the larger the physical property value is above a predetermined lower limit, the higher the reward for the Monte Carlo tree search. Alternatively, when designing a functional molecule with better physical properties, the lower the physical property value is, for example, when designing a molecule with a low dielectric constant, it is more preferable to set the calculated physical property value so that the smaller the physical property value is below a predetermined upper limit, the higher the reward for the Monte Carlo tree search. By setting up the machine learning in this way, it is possible to design novel functional molecules that exceed the limitations of molecules included in the training data. In the following embodiments, machine learning using Monte Carlo tree search in combination with a recurrent neural network and density functional theory, etc., will be described in detail as a method for generating design data for organic molecules with long-wavelength light absorption.

[0051] In the Monte Carlo tree search simulation procedure, it is preferable to use the trained recurrent neural network to generate a string of the final terminal node and generate a complete molecular string. For example, for the generated molecule (I), the optical absorption wavelength is calculated by TD-DFT calculation, and the reward r(I) of this molecule (I) is calculated using the following equations (3) and (4) (e.g., S33, S34, S35 in Figure 3).

[0052]

number

[0053] (In the above formula (3), v is the physical property value calculated by density functional theory, SA is the SA score (synthetic accessibility score), which is an actual value of 1-10 calculated from the structure of each molecule according to the method described in the following non-patent document. θ is the standard for the physical property value to be compared, and F(v) is 0.5 when v = θ, and is greater than 0.5 when v is greater than θ, and is 1 when v is infinity.) (Non-patent literature: Ertl, P., Schuffenhauer, A. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. J Cheminform. 1, 8 (2009).)

[0054] In the backpropagation procedure, it is preferable to propagate the reward r(I) of the molecule (I) toward the root node, and reflect the reward information in the Monte Carlo tree (FIG. 3, S51).

[0055] <Second generation step (S12)> In the second generation step (S12), the first generation data obtained in the first generation step (S11) includes molecular structure data and physical property data obtained by quantum chemical calculations, and the functional molecules having certain physical property values ​​are screened from the first generation data to generate second generation data (e.g., FIG. 3, S36 to S38). When designing a functional molecule with better physical properties as the physical property value increases, molecules with a predetermined physical property value or higher are screened to generate second generated data. For example, when designing an organic molecule with a long wavelength (described below), organic molecules with a predetermined optical absorption wavelength or higher are screened to generate second generated data. Alternatively, when designing a functional molecule with better physical properties as the physical property value decreases, molecules with a predetermined physical property value or lower are screened to generate second generated data. For example, when designing a molecule with a low dielectric constant, molecules with a predetermined dielectric constant or lower are screened to generate second generated data.

[0056] In the second generation step (S12), the physical properties of the first generation data may be recalculated using only the molecular structure of the first generation data obtained from the first generation step (S11) using a simulation method different from the quantum chemical calculation used in the first generation step (S11) or the same quantum chemical calculation but with different calculation accuracy (e.g., different basis functions).The recalculated physical properties may then be used to screen the first generation data for functional molecules having certain physical properties to generate second generation data.This allows for both high-speed first generation step (S11) and high accuracy in the second generation step (S12).

[0057] In the second generation step (S12), the molecular structure and physical property data of the first generation data obtained in the first generation step (S11) may be used to first screen the functional molecules having a predetermined first physical property value to generate intermediate data, and the physical property values ​​of the molecules in the intermediate data may be recalculated using a simulation method different from the quantum chemical calculation used in the first generation step (S11) or the same quantum chemical calculation but with a different calculation accuracy (e.g., a different basis function).The recalculated physical property values ​​may then be used to screen the functional molecules having a predetermined second physical property value from the intermediate data to generate second generation data.

[0058] <Substructure selection step (S13)> The "substructure" refers to a partial structure contained in the molecule that particularly contributes to the physical property. The substructure does not need to be a linked structure in the molecular structure (two-dimensional planar structure or three-dimensional steric structure) of the molecule, and may consist of two or more structures separated by a predetermined distance. In the embodiment described below, when training data expressed in SMILES notation is used, for example, a chromophore that contributes to long-wavelength light absorption may be used as the substructure, and a character string representing the chromophore in SMILES notation may be used as the substructure.

[0059] One or more of the partial structures may be selected from the partial structure candidate list. There are no particular limitations on the method of selection, as long as a partial structure that contributes to the above physical properties is selected. For example, the following three methods can be mentioned. [1] A method of selecting a partial structure based on the physical properties and common technical knowledge about the molecules included in the training data; [2] A method for automatically extracting and selecting possible substructures by analyzing the molecules and their physical properties contained in the training data; [3] A method of using a database of partial structures and selecting those contained in the database that contribute to the above physical properties. In the examples described below, the eight partial structures shown in Table 1 were selected from the viewpoint of contributing to long wavelength light absorption, and a partial structure candidate list for the embodiment was created.

[0060] For the selected partial structure candidates, the respective proportions are calculated in the training data and the second generated data obtained in the first generation step (S11), and the enrichment of each partial structure is calculated using the formula (1B). For example, if the number of molecules included in the training data is Nt and the number of molecules containing substructure f is Ntf, then the proportion of molecules containing substructure f in the training data, Pt(f) = Ntf / Nt. Similarly, if the number of molecules included in the second generated data is Nd and the number of molecules containing substructure f is Ndf, then the proportion of molecules containing substructure f in the first generated data, Pd(f) = Ndf / Nd. Using formula (1B), Pd(f) / Pt(f) is calculated as the enrichment of substructure f.

[0061] In the second generated data, a substructure is selected using enrichment. A substructure with a high enrichment can be selected. This is because molecules having a substructure with a high enrichment are particularly important molecules among the molecules generated by machine learning. For example, substructures with the first to third highest enrichments can be selected in descending order of enrichment. It is preferable to select the substructure with the highest enrichment. In the examples described below, the substructure with the highest enrichment was selected from the substructure candidate list in Table 1 of the examples (for example, FIG. 3, S39).

[0062] <Third generation step (S14)> Then, among the molecules included in the second generation data, molecules containing the substructure selected in the substructure selection step (S13) are screened to generate design data for functional molecules having predetermined physical properties (e.g., Figure 3, S40). In the examples described below, organic molecule design data was generated that included 22 molecules with the most highly enriched partial structures from the molecules included in the second generated data.

[0063] <Optimal design step> The method for generating design data of a functional molecule of this embodiment may further include an optimum design step of designing an optimum molecule from the design data. The optimization design step may, for example, involve re-evaluating the design data obtained from the third generation step (S14) using a quantum chemistry calculation method different from that used in the previous step to design an optimal molecule, or re-evaluating the design data using the same quantum chemistry calculation method as that used in the previous step but with a higher accuracy (requiring more computational time). The optimization design step may, for example, involve synthesizing a molecule contained in the second generation data obtained from the second generation step (S12) or a molecule having a similar structure, and evaluating the physical properties of the synthesized molecule. A molecule having a similar structure to a molecule contained in the design data is a molecule that, from the perspective of the structure that contributes to the physical properties, has the same basic structure as the molecule contained in the design data, but has a structure obtained by removing or replacing a partial structure that is difficult to synthesize with a partial structure that is easy to synthesize.

[0064] (Functional molecule design program) The functional molecule design program of this embodiment is a program for causing a computer to execute a method for generating design data for functional molecules having physical properties that can be calculated by quantum chemical calculations. A first generation step (S11) of generating first generated data by machine learning using training data; a second generation step (S12) of generating second generation data by screening the functional molecule from the first generation data using quantum chemical calculations; a substructure selection step (S13) of selecting a substructure that is a part of the functional molecule from the second generated data using the enrichment of the substructure; a third generation step (S14) of screening functional molecules having the partial structure from the second generation data to generate design data; Includes. The enrichment degree of the partial structure is a value calculated by the following formula (1B). Enrichment = fraction of molecules containing the substructure in the second generation data / fraction of molecules containing the substructure in the training data (1B) The physical property is preferably at least one selected from the group consisting of absorption wavelength, dielectric properties, oxidation-reduction potential, polarization properties, ionic conductivity, and cohesive energy.

[0065] In the method for generating design data of a functional molecule described above, a preferred embodiment of the method for generating the design data is described. From another viewpoint, the design program of this embodiment is a program for causing a computer to execute each step of the preferred embodiment of the method for generating the design data.

[0066] (Computer-readable recording medium recording a program for designing functional molecules) The present embodiment also includes a computer-readable recording medium on which the above-described functional molecule design program and preferred embodiments are recorded. Examples of the recording medium include tapes such as magnetic tapes and cassette tapes, disks including magnetic disks such as floppy disks (registered trademark) and hard disks, and optical disks such as CD-ROMs, MOs, MDs, DVDs, and CD-Rs, cards such as IC cards (including memory cards) and optical cards, and semiconductor memories such as mask ROMs, EPROMs, EEPROMs (registered trademark), and flash ROMs.

[0067] {Third embodiment} (Method for generating design data for organic molecules with long wavelength light absorption ability) Hereinafter, as a third embodiment, a method for generating design data of an organic molecule having long wavelength light absorption ability (hereinafter, sometimes referred to as the method for generating design data of this embodiment) will be described as a method for generating design data of a functional molecule. The method for generating design data of this embodiment includes the following first generation step, second generation step, substructure selection step, and third generation step. First generation step: A step of generating first generation data using machine learning using training data. Second generation step: generating second generation data by screening the organic molecules using quantum chemical calculations from the first generation data. Partial structure selection step: A step of selecting a chromophore structure from the second generated data using the enrichment of the chromophore structure. Third generation step: A step of selecting molecules having the chromophore structure from the second generation data to generate design data. The enrichment degree is expressed by the following formula (5).

[0068] Enrichment = fraction of molecules containing chromophore structures in the second generation data / fraction of molecules containing chromophore structures in the training data (5)

[0069] <First generation step> The quantum chemical calculations for generating design data in this embodiment use TD-DFT to obtain the electronic state of organic molecules and calculate the optical absorption wavelength. For functionals and basis functions, B3LYP / 6-311+G*, etc., are used.

[0070] In the method for generating design data according to this embodiment, molecules are selected from a molecular database written in a linear notation, such as SMILES notation. In this case, molecules may be selected randomly, or important molecules may be selected. Alternatively, the elements to be used may be narrowed down and selected. Examples of molecular databases using SMILES notation include ZINC15 and PUBCHEM. When using a molecular database that does not use a linear notation, training data can be generated by converting the notation to a linear notation, such as SMILES notation. For example, in the examples described below, a molecular database called ZINC15 written in SMILES notation is used, and molecules are randomly selected with the elements limited to carbon (C), nitrogen (N), and oxygen (O). The element type restrictions can be freely relaxed, allowing the use of a wider variety of element types, enabling the design of more interesting functional molecules.

[0071] The machine learning related to the method for generating design data in this embodiment uses reinforcement learning such as Monte Carlo tree search, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations such as density functional theory.

[0072] In the machine learning method for generating design data according to this embodiment, a recurrent neural network is first trained using the training data. For example, when a character string of a substructure written in SMILES notation is input, the trained recurrent neural network automatically generates the next character to the right of it, and by repeating this process, can finally generate a character string of the complete molecule written in SMILES notation. Using this trained recurrent neural network, rollout is performed in the simulation of Step 3 of the Monte Carlo tree search described below.

[0073] In the machine learning according to the method for generating design data of this embodiment, the numerator of the first generated data is generated using a Monte Carlo tree search. As shown in FIG. 4, the Monte Carlo tree search includes a selection step 1, an expansion step 2, a simulation step 3, and a backpropagation step 4.

[0074] In the selection procedure 1, the score u of the child node of the parent node is calculated using the following formula (2): i is preferably calculated using the Upper Confidence Bound 1 (UCB1) algorithm.

[0075]

number

[0076] (In equation (2), C is the search parameter, and w i / v i is the expected reward of the child node, and w iis the total reward, and v i is the number of visits to the child node, and V parent is the number of visits to the parent node.) The calculated score u i Select a child node based on the

[0077] In the expansion step 2, the selected child node is expanded.

[0078] In the simulation procedure 3, the trained recurrent neural network is used to generate a string of the final terminal node and generate a complete molecular string. For the generated molecule (I), it is preferable to calculate the optical absorption wavelength by TD-DFT calculation and calculate the reward r(I) for this molecule (I) using the following equations (3) and (4).

[0079]

number

[0080] (In the above formulas (3) and (4), v is the light absorption wavelength (unit: nm) calculated by density functional theory, SA is the SA score (synthetic accessibility score), θ is the reference wavelength for comparison, and F(v) is 0.5 when v = θ, a value greater than 0.5 when v is greater than θ, and 1 when v is infinity. In the examples described below, θ was set to 400 nm.)

[0081] In the backpropagation step 4, the reward r(I) of the numerator (I) is propagated toward the root node, and the reward information is reflected in the Monte Carlo tree. At this time, the parameters are updated according to the following equations (2A) and (2B).

[0082] w i C = w i p + r(I) (2A) v i c = v i p+1 (2B)

[0083] (In equations (2A) and (2B), w i C , w i p are the sums of the current and previous rewards, respectively; v i c , v i p is the number of visits to the child node this time (Current) and last time (Previous).

[0084] <Second generation step> In the second generation step, the first generation data obtained from the first generation step includes molecular structure data and optical absorption wavelength data obtained by quantum chemical calculation, and the organic molecules having optical absorption wavelengths equal to or greater than a certain wavelength are screened from the first generation data to generate second generation data.

[0085] In the second generation step, the optical absorption wavelength of the first generation data may be recalculated using only the molecular structure of the first generation data obtained from the first generation step, using a simulation method different from that used in the quantum chemical calculation or the same quantum chemical calculation but with different calculation accuracy (e.g., different basis functions).The recalculated optical absorption wavelength may then be used to screen the first generation data for organic molecules having wavelengths equal to or greater than a certain optical absorption wavelength, thereby generating second generation data.This allows for both a high-speed first generation step and high accuracy in the second generation step.

[0086] In the second generation step, intermediate data may be generated by first screening organic molecules having wavelengths equal to or greater than a predetermined first optical absorption wavelength using the molecular structure and optical absorption wavelength of the first generation data obtained in the first generation step, and then recalculating the organic molecules of the molecules in the intermediate data using a simulation method different from that used in the quantum chemical calculation in the first generation step or the same quantum chemical calculation with different calculation accuracy (e.g., different basis functions).The recalculated physical property values ​​may then be used to screen the organic molecules having wavelengths equal to or greater than a predetermined second optical absorption wavelength from the intermediate data to generate second generation data.

[0087] <Substructure selection step (S13)>

[0088] In the partial structure selection step (S13) of the method for generating design data according to this embodiment, chromophore structures that contribute to long-wavelength light absorption are considered as partial structures. Examples of these chromophore structures include ketones, quinones, and naphthoquinones. Specific examples include the eight chromophore structures listed in Table 1 described in the Examples. The chromophore structure may be automatically extracted and selected by analyzing the molecules included in the training data and their light absorption wavelengths. Alternatively, a database of chromophore structures may be used to select from the database those that contribute to long wavelength light absorption. Alternatively, a database relating to chromophore structures may be used, and the entire database may be used.

[0089] For the selected chromophore structure, the respective proportions are calculated in the training data and the second generated data obtained in the second generation step, and the enrichment of each chromophore structure is calculated using the formula (1B). For example, if the number of molecules included in the training data is Nt and the number of molecules containing chromophore structure f is Ntf, then the proportion of molecules containing chromophore structure f in the training data, Pt(f) = Ntf / Nt. Similarly, if the number of molecules included in the first generated data is Nd and the number of molecules containing chromophore structure f is Ndf, then the proportion of molecules containing chromophore structure f in the first generated data, Pd(f) = Ndf / Nd. Using formula (1B), Pd(f) / Pt(f) is calculated as the enrichment of chromophore structure f. For example, as shown in Table 1 described in the Examples, the enrichment of each chromophore structure f was calculated.

[0090] In the second generated data, a chromophore structure is selected using the enrichment. A chromophore structure with a high enrichment can be selected. This is because molecules having a chromophore structure with a high enrichment are particularly important molecules among the molecules generated by machine learning. For example, chromophore structures with the first to third highest enrichments can be selected in descending order of enrichment. In this example, the chromophore structure of 1,2-naphthoquinone, which has the highest enrichment, was selected.

[0091] <Third generation step> It is preferable to generate design data by screening molecules having the chromophore structure with the highest enrichment among the molecules included in the second generated data. For example, in this example, design data including 22 molecules having the chromophore structure of 1,2-naphthoquinone, which showed the highest enrichment of 78.5 among the molecules included in the second generated data, was generated.

[0092] <Optimal design step> The method of generating design data for an organic molecule capable of absorbing long-wavelength light according to this embodiment may further include an optimum design step of designing an optimum molecule from the obtained design data. In the optimization design step, to design an optimal molecule, the molecules included in the design data obtained from the third generation step, such as a molecule having the chromophore structure of 1,2-naphthoquinone in the example, are compared using the quantum chemical calculation values ​​of the optical absorption wavelengths obtained by the machine learning in the first generation step. Among these, the molecule with the longest wavelength is selected, and the possibility of synthesis is considered. Alternatively, a molecule having the same basic molecular structure as the molecule with the longest wavelength and being easy to synthesize is redesigned, and its optical absorption wavelength is calculated using the same quantum chemical calculation method as above, to confirm that the long wavelength absorption is maintained. Then, the molecule with the longest wavelength included in the design data obtained in the third generation step, or a molecule having a structure similar thereto, may be synthesized, and the physical properties of the synthesized molecule may be evaluated. A molecule having a structure similar to that of a molecule included in the design data is a molecule having the same basic structure as the molecule included in the design data, from the perspective of a chromophore that contributes to long-wavelength light absorption, with partial structures that are difficult to synthesize removed or replaced with partial structures that are easy to synthesize.

[0093] (program) The program for generating design data for organic molecules capable of absorbing long-wavelength light in this embodiment is a program for causing a computer to execute the following first generation step, second generation step, partial structure selection step, and third generation step. First generation step: A step of generating first generation data using machine learning using training data. Second generation step: A step of generating second generation data by screening the organic molecules using quantum chemical calculations from the first generation data. Partial structure selection step: A step of selecting a chromophore structure from the second generated data using the enrichment of the chromophore structure. Third generation step: A step of selecting molecules having the chromophore structure from the second generation data to generate design data. The enrichment degree is expressed by the following formula (1C).

[0094] Enrichment = fraction of molecules containing chromophore structures in the second generation data / fraction of molecules containing chromophore structures in the training data (1C)

[0095] Regarding the method for generating design data for organic molecules having long-wavelength light absorption ability, a preferred embodiment of the method for generating the design data is described. From another viewpoint, the design program of this embodiment is a program for causing a computer to execute each of the above steps.

[0096] (Computer-readable recording medium on which a program is recorded) The present embodiment also includes a computer-readable recording medium having recorded thereon a program for generating design data for organic molecules having long-wavelength light absorption capability, as well as preferred embodiments thereof. Examples of the recording medium include tape-based media such as magnetic tape and cassette tape, disk-based media including magnetic disks such as floppy disks (registered trademark) and hard disks, and optical disks such as CD-ROMs, MOs, MDs, DVDs, and CD-Rs, card-based media such as IC cards (including memory cards) and optical cards, and semiconductor memory-based media such as mask ROMs, EPROMs, EEPROMs (registered trademark), and flash ROMs.

[0097] (Molecules included in the design data and their synthesis studies) The design data obtained by the method for generating design data for functional molecules includes molecules having the above-mentioned physical properties. The physical properties of these designed molecules may be evaluated using other known quantum chemical calculation methods or experimental methods. Furthermore, methods for synthesizing these designed molecules may be investigated. The design data obtained by the method for generating design data for organic molecules capable of absorbing long-wavelength light according to this embodiment includes design data for molecules having a specific chromophore structure. The light absorption wavelength of these designed molecules may be evaluated using other known quantum chemical methods or experimental methods. Furthermore, synthesis methods for these designed molecules may be investigated.

[0098] (Optimal molecules and their synthesis) In a preferred embodiment of the method for generating design data for a functional molecule of this embodiment, when it further includes an optimum design step, the optimum molecule obtained in the optimum design step is actually synthesized. If the method of generating design data for organic molecules having long wavelength light absorption ability according to this embodiment further includes an optimum design step, the optimum molecule obtained in the optimum design step is actually synthesized. Specifically, the molecule represented by the following formula (6) and its derivatives are exemplified. Examples of derivatives of the molecule represented by the following formula (6) include molecules in which a total of 10 hydrogen atoms, namely, the five hydrogen atoms on the phenyl group on the right side of the following formula (6), the four hydrogen atoms on the benzene ring on the left side of the following formula (6), and the hydrogen atom bonded to the carbon at the 4-position of the 1,2-naphthoquinone chromophore, are substituted with 1 to 10 substituents. In the case of two or more substituents, these may be the same or different. The number of these substitutions is preferably five or less, more preferably three or less, and most preferably one. The type of the substituent is not particularly limited as long as it does not affect the absorption characteristics of the 1,2-naphthoquinone chromophore, and examples thereof include alkyl groups such as methyl and ethyl groups.

[0099] [ka] [Example]

[0100] The present invention will be described in detail below with reference to examples, but the present invention is not limited to these examples. Example 1 "Design of organic molecules with long wavelength light absorption ability" <First generation step> The training data used in this example was a molecular database called ZINC15, written in SMILES notation. ZINC15 is a free molecular database consisting of over 750 million molecules and provides virtual screening. Specifically, for the training data used in this example, 153,253 molecules were randomly selected from ZINC15, with the available elements limited to C, N, and O. The method for generating the first generated data in this example was the same as the first generation step described in the "Method for generating design data for organic molecules having long wavelength light absorption ability" in the third embodiment. As a result, data for 45,321 molecules was generated. In the first generation step of this embodiment, the software, libraries, and respective calculation times and calculation environments used are as follows. Software and libraries: Python, RDkit, keras, Gaussian16 Computation time (RNN learning): 12 hours Computing environment (RNN learning): NVIDIA Tesla V100 Computation time (molecule generation): 120 hours Computation environment (molecule generation): Intel Xeon 2048 core

[0101] <Second generation step> The method for generating the second generated data in this example was the same as the method for the second generation step described in the third embodiment. From the 45,321 molecules, 380 molecules with an optical absorption wavelength of 800 nm or more calculated by TD-DFT and an oscillator strength of 0.1 or more were selected to generate the second generated data in this example.

[0102] <Substructure selection step> "Evaluation of Enrichment Degree" The enrichment of the molecules in the generated second-generation data was evaluated to determine which chromophore structures are suitable for long-wavelength light absorption. The enrichment was defined as the ratio of a specific chromophore structure in the training data to its value in the generated second-generation data and was calculated using the above formula (1C). As a result, among molecules exhibiting long-wavelength light absorption, quinone structures such as 1,2-quinone, 1,4-quinone, and anthraquinone were more enriched than ketones and diketones (Table 1). Molecules with 1,4-quinone and anthraquinone structures have been used as dyes since the BC era and were produced industrially in the mid-1800s. Among them, 1,2-naphthoquinone exhibited higher enrichment than 1,2-naphthoquinone and anthraquinone, demonstrating long-wavelength light absorption. These results indicated that machine learning predicted that 1,2-naphthoquinone is an important factor in long-wavelength light absorption. Dyes with anthraquinone and 1,2-naphthoquinone structures are well known. However, while 1,2-naphthoquinone is known as a cofactor and is used as a building block in heterocyclic synthesis, it has received little attention as a dye, making this an important discovery through machine learning.

[0103] [Table 1]

[0104] The general formula of the molecule having each chromophore structure in Table 1 is shown in formula (7) below. [ka]

[0105] <Third generation step> From the 380 pieces of second generation data, 22 different molecules represented by the following formulas (5-1) to (5-22) containing the selected partial structure (chromophore) 1,2-naphthoquinone structure were selected to generate the design data for this example.

[0106] [ka]

[0107] <Design data molecules and their synthesis methods> Among the 22 molecules (5-1) to (5-22) designed by the formulas (5-1) to (5-22) above, the synthesis methods of molecules (5-1) to (5-5), (5-7), (5-9) to (5-10), (5-12) to (5-15), and (5-18) to (5-22) were investigated by retrosynthetic analysis. Molecule (5-17) was actually synthesized in Synthesis Examples 1 and 2 above.

[0108] The synthesis of molecule (5-1) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-1) can be synthesized. [ka]

[0109] The synthesis of molecule (5-2) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-2) can be synthesized. [ka]

[0110] The synthesis of molecule (5-3) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-3) can be synthesized. [ka]

[0111] The synthesis of molecule (5-4) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-4) can be synthesized. [ka]

[0112] The synthesis of molecule (5-5) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-5) can be synthesized. [ka]

[0113] The synthesis of molecule (5-7) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-7) can be synthesized. [ka]

[0114] The synthesis of molecule (5-9) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-9) can be synthesized. [ka]

[0115] The synthesis of molecule (5-10) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-10) can be synthesized. [ka]

[0116] The synthesis of molecule (5-12) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-12) can be synthesized. [ka]

[0117] The synthesis of molecule (5-13) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-13) can be synthesized. [ka]

[0118] The synthesis of molecule (5-14) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-14) can be synthesized. [ka]

[0119] The molecule (5-15) is commercially available. [ka]

[0120] The synthesis of molecule (5-18) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-18) can be synthesized. [ka]

[0121] The synthesis of molecule (5-19) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-19) can be synthesized. [ka]

[0122] The synthesis of molecule (5-20) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-20) can be synthesized. [ka]

[0123] The synthesis of molecule (5-21) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-21) can be synthesized. [ka]

[0124] The synthesis of molecule (5-22) is described according to the following retrosynthetic analytical equation: It has been found that molecule (5-22) can be synthesized. [ka]

[0125] <Optimal molecule> Among the organic molecules represented by formulas (5-1) to (5-22), we focused on molecule (5-17) (molecule 1a in Figure 5), which exhibits the longest wavelength absorption among these organic molecules. As shown in Figure 5, molecule 1a has a structure consisting of an enol and a carbonyl skeleton bonded to a 1,2-naphthoquinone skeleton, and TD-DFT calculations predicted an optical absorption wavelength of 941 nm (Figure 5). As the actual synthetic target, we selected a molecule with simplified functional groups that do not affect the optical absorption wavelength. Since the carbonyl skeleton does not have a significant effect on the optical absorption wavelength, it was replaced with hydrogen (molecule 1b in Figure 5). From the TD-DFT calculations, the optical absorption wavelength of molecule 1b is predicted to be 941 nm, which is the same as the optical absorption wavelength of molecule 1a (λ 1b = 835 nm). Because it was difficult to introduce a triazole group, the triazole group in molecule 1b was also replaced with a phenyl group with a π-conjugated structure (Figure 5, molecule 1c).

[0126] TD-DFT calculations revealed that the optical absorption wavelength of molecule 1c was 789 nm, which was shorter than that of molecules 1a and 1b.

[0127] On the other hand, because the enol skeleton is difficult to isolate chemically, we selected molecule 1d, in which the enol skeleton of molecule 1c was replaced with a simple olefin, as the synthetic target (Figure 5).

[0128] The retrosynthetic analytical formula for molecule 1d (molecule represented by formula (5-22)) is shown in Figure 6. Molecule 1d can be synthesized by oxidation of molecule 2, and we thought that 2 could be synthesized using the cross-coupling reaction of olefins with naphthols having leaving groups and metal groups.

[0129] Starting with molecule 3, the Suzuki-Miyaura coupling reaction with vinylboronic acid pinacol ester afforded molecule 2, in which an olefin moiety was introduced into the naphthol skeleton (Figure 7: Scheme 1). Molecule 2 was oxidized with 2-iodoxybenzoic acid (IBX) to give the target molecule 1b as a dark purple solid in 95% yield. Molecule 1d was air-stable and soluble in many common solvents, such as CHCl3, CHCl2, THF, CH3CN, and acetone. The product was characterized by various NMR spectroscopy techniques, as described in detail in Examples 1 and 2 below. A solution of 1d in CH3CN (0.1 mmol / L) exhibited a reddish-purple color and showed a longest wavelength absorption at 550 nm. TD-DFT calculations using APFD / 6-311++G** showed a spectrum close to the experimental one, with an absorption at 600 nm. The results are shown in Figure 8.

[0130] The synthesis of molecule 1d will be described in detail below in Synthesis Examples 1 and 2. Materials used: Unless otherwise specified, the materials used in the following synthesis examples were commercially available chemicals and were used as received.

[0131] (Analysis method) [ 1 H-NMR Measurement conditions: (a) Solvent: deuterated chloroform (CDCl3) (b) Internal standard: tetramethylsilane (TMS, 0 ppm) (c) Measurement temperature: 25 ℃ (d) Resonance frequency: 400MHz (e) Measuring device: JEOL ECS-400

[0132] [ 13 C-NMR Measurement conditions: (a) Solvent: deuterated chloroform (CDCl3) (b) Internal standard: deuterated chloroform (CDCl3, 77 ppm) (c) Measurement temperature: 25 ℃ (d) Resonance frequency: 100 MHz (e) Measuring device: JEOL ECS-400

[0133] [UV / vis spectrum] (a) Solvent: deuterated chloroform (CDCl3) (b) Measurement temperature: 25 ℃ (c) Atmosphere: Air (d) Measuring device: JASCO V670

[0134] (Synthesis Example 1) <Synthesis of molecule 3 in Figure 7> A solution of 3-bromo-2-naphthol (1.0 g, 4.5 mmol), trans-2-phenylvinylboronic acid pinacol ester (1.6 g, 6.8 mmol), Pd(OAc)2 (80.8 mg, 0.36 mmol), dppf (0.4 g, 0.72 mmol), potassium phosphate tripotassium (4.8 g, 22.5 mmol), and anhydrous dioxane (20 mL) was heated at 110 °C for 24 h under a nitrogen atmosphere. The reaction mixture was passed through a pad of silica gel with dichloromethane as the eluent and concentrated under reduced pressure to give the crude mixture. Purification by silica gel chromatography (dichloromethane / hexane = 8 / 2) afforded molecule 3 (601.9 mg, 54%). Under the above conditions, 1 H-NMR and 13 C-NMR was measured, and the results are shown below.

[0135] 1 H-NMR (CDCl3, 400 MHz): δ = 5.15 (s, 1H, OH), 7.14 (s, 1H, ArH), 7.29-7.42 (m, 5H, ArH), 7.32 (d, 1H, J = 16.5 Hz, CH), 7.39 (s, 1H, ArH), 7.51 (d, 1H, J = 16.5 Hz, CH), 7.59 (d, 1H, 7.3 Hz, ArH), 7.59 (d, 1H, 8.2 Hz, ArH), 7.65 (d, 1H, 8.2 Hz, ArH), 7.79 (d, 1H, 7.3 Hz, ArH), 8.01(s, 1H, ArH); 13C-NMR (CDCl3, 100 MHz): δ = 110.1, 123.4, 124.1, 126.0, 126.5, 126.7, 126.8, 127.0, 127.9, 128.0, 128.8, 129.3, 131.4, 134.0, 137.7, 151.7.

[0136] (Synthesis Example 2) <Synthesis of molecule 1d in Figure 7> A solution of the above molecule 3 (200.0 mg, 0.81 mmol) obtained in Synthesis Example 1, IBX (o-iodoxy benzoic acid, purity >39%) (640.0 mg, 0.89 mmol), and DMF (N,N-dimethylformamide) (5 mL) was stirred at room temperature for 1 hour. The reaction mixture was quenched with distilled water and extracted with ethyl acetate. The combined organic layer was washed with saturated aqueous NaCl solution, dried over MgSO4, filtered, and concentrated under reduced pressure to obtain a crude mixture. Purification by silica gel chromatography (dichloromethane / hexane = 9 / 1) yielded molecule 1d (200.8 mg, 95%). Under the above measurement conditions, 1 H-NMR and 13 C-NMR was measured, and the results are shown below. 1 H-NMR (CDCl3, 400 MHz): δ =7.05 (d, 1H, J = 16.5 Hz, CH), 7.30 (m, 1H, ArH), 7.37(m, 1H, ArH), 7.40 (m, 2H, ArH), 7.45 (td, 1H, J = 7.6, 1.4 Hz, ArH), 7.47 (d, 1H, J = 16.5 Hz, CH), 7.49 (s, 1H, CH), 7.54 (d, 2H, 7.3 Hz, ArH), 7.64 (td, 1H, J = 7.6, 1.4 Hz, ArH) 8.07 (dd, 1H, J = 7.6, 1.4 Hz, ArH); 13C-NMR (CDCl3, 100 MHz): δ = 121.1, 127.2, 128.8, 128.9, 130.0, 130.3, 130.6, 133.8, 135.1, 135.8, 136.2, 136.8, 138.8 (Consideration) Until now, chemists have synthesized functional molecules by optimizing the properties of molecules discovered in nature or synthesized by chance. However, the proposed AI technology (machine learning) allows computers to search for molecules with desired properties, select highly enriched substructures, and design molecules that incorporate these substructures. This could potentially lead to the discovery of molecules that have never received attention before, or even molecules that chemists have never considered. In the third embodiment, a method for generating organic molecules with long-wavelength light absorption capabilities, combining Monte Carlo tree search, recurrent neural networks, and density functional theory, produced highly enriched 1,2-naphthoquinone, a chromophore that has received little attention. Furthermore, the structure of an organic molecule containing the 1,2-naphthoquinone chromophore generated by this method was optimized to synthesize a new red-purple molecule with absorption at 550 nm. [Industrial Applicability]

[0137] According to this embodiment, it is possible to provide a method for generating design data for functional molecules and a program for designing functional molecules, which can generate molecules by effectively utilizing machine learning and quantum chemical calculations and design promising functional molecules by focusing on the enrichment of partial structures that contribute to predetermined physical properties. It is also possible to provide functional molecules obtained by the method for generating design data for functional molecules. This embodiment is expected to accelerate the development of functional molecules in the field of organic electronics, such as light-harvesting materials for solar cells, electricity storage materials, and light-emitting / host materials for organic electroluminescence (EL).

Claims

1. A method for generating design data for an object having properties that can be predicted by computer simulation, comprising: a first generation step of using generated data obtained by machine learning from training data including data on the configuration of an object having the property and data on the value of the property of the object as first generated data; a second generation step of generating second generation data by screening the object having the property from the first generation data using the property value calculated by computer simulation; a substructure selection step of selecting a substructure from the second generated data based on the enrichment of the substructure of the object; a third generation step of screening the object having the substructure selected in the substructure selection step from the second generation data to generate design data of the object; Including, In the first generating step, The machine learning is reinforcement learning using a Monte Carlo tree search algorithm, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations; the first generated data obtained by the machine learning includes configuration data of an object and value data of properties of the object obtained by the computer simulation, A design data generating method, characterized in that the enrichment degree of the partial structure is a value calculated by the following formula (1A). Enrichment = Proportion of items containing the partial structure in the second generated data / Proportion of items containing the partial structure in the training data (1A)

2. A method for generating design data for a functional molecule having physical properties that can be calculated by quantum chemical calculation, comprising: a first generation step of using generated data obtained by machine learning from training data including data on the chemical structure of a functional molecule having the physical property and data on the value of the physical property of the functional molecule as first generated data; a second generation step of generating second generation data by screening the functional molecules using the values ​​of the physical properties calculated by quantum chemical calculation from the first generation data; a substructure selection step of selecting the substructure from the second generated data using the enrichment of the substructure of the functional molecule; a third generation step of screening functional molecules having the partial structure from the second generation data to generate design data; Including, In the first generating step, The machine learning is reinforcement learning using a Monte Carlo tree search algorithm, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations; the first generated data obtained by the machine learning includes chemical structure data of a functional molecule and physical property value data of the functional molecule obtained by the quantum chemical calculation, A method for generating design data of a functional molecule, wherein the enrichment of the partial structure is a value calculated by the following formula (1B): Enrichment = Proportion of molecules containing the substructure in the second generation data / Proportion of molecules containing the substructure in the training data (1B)

3. the physical property is a light absorption wavelength, 3. The method for generating design data for a functional molecule according to claim 2, wherein the functional molecule is an organic molecule having a long wavelength light absorption ability.

4. the Monte Carlo tree search includes a selection step, an expansion step, a simulation step, and a backpropagation step; In the selection procedure, the score u of the child node of the parent node is calculated using the following formula (2): ｉ is calculated using the UCB1 algorithm, [Equation 1] (In equation (2), C is a search parameter, and w ｉ / v ｉ is the expected reward of the child node, and w ｉ is the total reward, and v ｉ is the visit count of the child node, and V ｐａｒｅｎｔ is the number of visits to the parent node.) Previous calculated score u ｉ Select the child node based on In the expansion step, expanding the selected child node; In the simulation procedure, a recurrent neural network is used to calculate the value of the physical property for a molecule represented by a child node expanded based on the rollout using TD-DFT, and a reward r(I) for this child node is calculated using the following equations (3) and (4): [Equation 2] (In the formulas (3) and (4), v is the physical property value calculated by TD-DFT, SA is the SA score of each molecule and ranges from 1 to 10, and θ is the standard for the physical property value to be compared. When v = θ, it becomes 0.5, when v is greater than θ, it becomes a value greater than 0.5, and when v is infinity, it becomes 1.) 3. The method for generating design data for functional molecules according to claim 2, wherein the reward r(I) is propagated toward the root node in the backpropagation procedure, and the reward information is reflected in a Monte Carlo tree.

5. A method for generating design data of a functional molecule described in any one of claims 2 to 4, wherein in the partial structure selection step, a partial structure having the highest enrichment in the enrichment degree of the partial structure is selected from the second generation data.

6. The method for generating design data for a functional molecule according to claim 5 , wherein the second generated data includes a molecule having a partial structure with the highest enrichment.

7. A method for generating design data of a functional molecule described in any one of claims 2 to 4, further comprising an optimization design step of designing a molecule having a similar structure to the molecule included in the design data obtained in the third generation step, actually synthesizing the molecule, and measuring the target physical properties.

8. A program for causing a computer to execute a method for generating design data for an object having properties that can be predicted by computer simulation, the method comprising: a first generation step of using generated data obtained by machine learning from training data including data on the configuration of an object having the property and data on the value of the property of the object as first generated data; a second generation step of generating second generation data by screening the object having the property from the first generation data using the property value calculated by computer simulation; a substructure selection step of selecting a substructure from the second generated data based on the enrichment of the substructure of the object; a third generation step of screening the object having the substructure selected in the substructure selection step from the second generation data to generate design data of the object; Including, In the first generating step, The machine learning is reinforcement learning using a Monte Carlo tree search algorithm, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations; the first generated data obtained by the machine learning includes configuration data of an object and value data of properties of the object obtained by the computer simulation, The program, wherein the enrichment of the partial structure is a value calculated using the following formula (1A): Enrichment = Proportion of items containing the partial structure in the second generated data / Proportion of items containing the partial structure in the training data (1A)

9. A program for causing a computer to execute a method for generating design data for a functional molecule having physical properties that can be calculated by quantum chemical calculation, the method comprising: a first generation step of using generated data obtained by machine learning from training data including data on the chemical structure of a functional molecule having the physical property and data on the value of the physical property of the functional molecule as first generated data; a second generation step of generating second generation data by screening the functional molecules using the values ​​of the physical properties calculated by quantum chemical calculation from the first generation data; a substructure selection step of selecting a substructure that is a part of the functional molecule from the second generated data using an enrichment degree of the substructure; a third generation step of screening functional molecules having the partial structure from the second generation data to generate design data; Including, In the first generating step, The machine learning is reinforcement learning using a Monte Carlo tree search algorithm, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations; the first generated data obtained by the machine learning includes chemical structure data of a functional molecule and physical property value data of the functional molecule obtained by the quantum chemical calculation, The program, wherein the enrichment of the partial structure is a value calculated using the following formula (1B): Enrichment = Proportion of molecules containing the substructure in the second generation data / Proportion of molecules containing the substructure in the training data (1B)

10. the physical property is a light absorption wavelength, The program according to claim 9 , wherein the functional molecule is an organic molecule capable of absorbing long-wavelength light.

11. A computer-readable recording medium having a program recorded thereon for causing a computer to execute a method for generating design data for an object having properties that can be predicted by computer simulation, the method comprising: a first generation step of using generated data obtained by machine learning from training data including data on the configuration of an object having the property and data on the value of the property of the object as first generated data; a second generation step of generating second generation data by screening the object having the property from the first generation data using the property value calculated by computer simulation; a substructure selection step of selecting a substructure from the second generated data based on the enrichment of the substructure of the object; a third generation step of screening the object having the substructure selected in the substructure selection step from the second generation data to generate design data of the object; Including, In the first generating step, The machine learning is reinforcement learning using a Monte Carlo tree search algorithm, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations; the first generated data obtained by the machine learning includes configuration data of an object and value data of properties of the object obtained by the computer simulation, The recording medium, wherein the concentration of the partial structure is a value calculated by the following formula (1A): Enrichment = Proportion of items containing the partial structure in the second generated data / Proportion of items containing the partial structure in the training data (1A)

12. A computer-readable recording medium storing a program for causing a computer to execute a method for generating design data for functional molecules having physical properties that can be calculated by quantum chemical calculation, the method comprising: a first generation step of using generated data obtained by machine learning from training data including data on the chemical structure of a functional molecule having the physical property and data on the value of the physical property of the functional molecule as first generated data; a second generation step of generating second generation data by screening the functional molecules using the values ​​of the physical properties calculated by quantum chemical calculation from the first generation data; a substructure selection step of selecting a substructure that is a part of the functional molecule from the second generated data using an enrichment degree of the substructure; a third generation step of screening functional molecules having the partial structure from the second generation data to generate design data; Including, In the first generating step, The machine learning is reinforcement learning using a Monte Carlo tree search algorithm, and in the Monte Carlo tree search, rollout is performed by combining a recurrent neural network and quantum chemical calculations; the first generated data obtained by the machine learning includes chemical structure data of a functional molecule and physical property value data of the functional molecule obtained by the quantum chemical calculation, The recording medium, wherein the enrichment degree of the partial structure is a value calculated by the following formula (1B): Enrichment = Proportion of molecules containing the substructure in the second generation data / Proportion of molecules containing the substructure in the training data (1B)

13. the physical property is a light absorption wavelength, The computer-readable recording medium storing the program according to claim 12, wherein the functional molecule is an organic molecule having a long wavelength light absorption ability.

Images